import warnings
warnings.filterwarnings('ignore')
# %matplotlib notebook

Soil moisture analysis switzerland

TOC

• Datasets
• Methods
• Analysis
  • Individual Stations
  • Overall

Comparison of modelled soil moisture (PREVAH, ERA5, ERA5-Land, Lisflood, UERRA) with in-situ soil moisture from the SwissSMEX (SWMX) dataset.

Datasets

Model data were interpolated to represent soil moisture at the same depth (50 cm) as the in situ measurement.

Models

• ERA5: DOI: 10.24381/cds.adbb2d47
• ERA5-Land: DOI: 10.24381/cds.e2161bac
• UERRA (MESCAN-SURFEX): access through CDS
• LISFLOOD: DOI: 10.23381/cds.3458969
• PREVAH: Zappa and Gurtz (2003) DOI: 10.1007/s10346-020-01400-y; Viviroli et al. (2009) DOI: 10.1016/j.envsoft.2009.004.001

PREVAH represents a high-resolution, locally optimised hydrological model and acts as a reference for this comparison.

Reference

SwissSMEX in-situ measurements: Mittelbach and Seneviratne (2012) DOI: 10.5194/hess-16-2169-2012

We selected 6 stations with long and uninterrupted time-series:

• Payerne
• Plaffeien
• Wynau
• Chamau
• Reckenholz
• Taenikon

import pandas as pd

swmx_sites = pd.read_csv('/mnt/CEPH_PROJECTS/ADO/SM/reference_data/SiteInfo_SwissSMEX0713_grassland_selected.csv',
                         skiprows=1)
print(swmx_sites.loc[:, ['Site', 'Elevation', 'Landuse', 'Texture', 'ID']])

         Site  Elevation    Landuse     Texture  ID
0     Payerne        490  grassland        loam   2
1   Plaffeien       1042  grassland  sandy loam   3
2       Wynau        422  grassland   silt loam   7
3      Chamau        400  grassland  sandy loam   8
4  Reckenholz        443  grassland        loam  10
5    Taenikon        536  grassland        loam  11

Methods

• SwissSMEX in-situ measurements were converted from mm to m^3/m^-3: $SM_{vol} = SM_{mm} / 500$
• To offset the impact of biases between the different models, the comparison is based on anomalies (SMA). We used to following equation to compute SMA's: $SMA = SM_t - \bar{SM} / \delta_{SM}$ where $SM_t$ is the current value of SM on any given data, $\bar{SM}$ is the long-term average, and $\delta_{SM}$ the related standard deviation. The reference period was selected for each site individually, based on the temporal overlap between all models.

Analysis

from model_comparison import initialize_stacks
ERA5, ERA5l, LISFLOOD, UERRA = initialize_stacks()

The following section shows several plots for each of the six selected SwissSMEX sites: time-series plots covering the full period of available data; time-series plots just for the years 2012 and 2018; Scatterplots which show a direct comparison between in-situ and modelled soil moisture.

The table summarizes the accuracy of modelled soil moisture, in terms of R and RMSE, with respect to the SwissSMEX measurements. The average in the last column refers to the average soil moisture computed over all stations.

R

--------	Payerne	Plaffeien	Wynau	Chamau	Reckenholz	Taenikon	Average
PREVAH	0.705	0.678	0.749	0.437	0.642	0.624	0.623
ERA5	0.617	0.592	0.680	0.442	0.614	0.574	0.647
ERA5-L	0.641	0.536	0.745	0.455	0.585	0.547	0.642
LISFLOOD	0.736	0.544	0.712	0.194	0.794	0.778	0.733
UERRA	0.721	0.678	0.634	0.541	0.539	0.657	0.577

RMSE

--------	Payerne	Plaffeien	Wynau	Chamau	Reckenholz	Taenikon	Average
PREVAH	0.803	0.842	0.760	1.164	0.908	0.906	0.914
ERA5	0.910	0.928	0.837	1.064	0.905	0.938	0.870
ERA5-L	0.865	1.005	0.749	1.059	0.934	0.964	0.878
LISFLOOD	0.752	0.859	0.775	1.274	0.670	0.662	0.741
UERRA	0.761	0.835	0.917	0.992	0.961	0.849	0.963

This first overview shows that correlations are relatively low in general. Interestingly, higher spatial resolution does not always lead to higher accuracies. Performances are relatively similar across the tested models.

Individual Stations

Payerne

from model_comparison import SwissSMEX_ts_create_plots

# setting
interval = 1
monthly=True
anomalies=True
year = [2012, 2018]

i_swmx = 'PAY'
anom_merged, merged = SwissSMEX_ts_create_plots(ERA5, ERA5l,
                                                     LISFLOOD,
                                                     UERRA,
                                                     swmx_sites.iloc[0],
                                                     interval,
                                                     year=year)

mindex = pd.MultiIndex.from_tuples([(i_swmx, i) for i in anom_merged.columns])
all_stations = merged.copy()
all_stations.columns = mindex

Plaffeien

i_swmx = 'PLA'
anom_merged, merged = SwissSMEX_ts_create_plots(ERA5, ERA5l,
                                        LISFLOOD,
                                        UERRA,
                                        swmx_sites.iloc[1],
                                        interval,
                                        year=year)

mindex = pd.MultiIndex.from_tuples([(i_swmx, i) for i in anom_merged.columns])

tmp = merged.copy()
tmp.columns = mindex
all_stations = pd.concat([tmp, all_stations], axis=1)

Wynau

i_swmx = 'WYN'
anom_merged, merged = SwissSMEX_ts_create_plots(ERA5, ERA5l,
                                        LISFLOOD,
                                        UERRA,
                                        swmx_sites.iloc[2],
                                        interval,
                                        year=year)

mindex = pd.MultiIndex.from_tuples([(i_swmx, i) for i in anom_merged.columns])

tmp = merged.copy()
tmp.columns = mindex
all_stations = pd.concat([tmp, all_stations], axis=1)

Chamau

i_swmx = 'CHM'
anom_merged, merged = SwissSMEX_ts_create_plots(ERA5, ERA5l,
                                        LISFLOOD,
                                        UERRA,
                                        swmx_sites.iloc[3],
                                        interval,
                                        year=year)

mindex = pd.MultiIndex.from_tuples([(i_swmx, i) for i in anom_merged.columns])

tmp = merged.copy()
tmp.columns = mindex
all_stations = pd.concat([tmp, all_stations], axis=1)

Reckenholz

i_swmx = 'REC'
anom_merged, merged = SwissSMEX_ts_create_plots(ERA5, ERA5l,
                                        LISFLOOD,
                                        UERRA,
                                        swmx_sites.iloc[4],
                                        interval,
                                        year=year)

mindex = pd.MultiIndex.from_tuples([(i_swmx, i) for i in anom_merged.columns])

tmp = merged.copy()
tmp.columns = mindex
all_stations = pd.concat([tmp, all_stations], axis=1)

Taenikon

i_swmx = 'TAE'
anom_merged, merged = SwissSMEX_ts_create_plots(ERA5, ERA5l,
                                        LISFLOOD,
                                        UERRA,
                                        swmx_sites.iloc[5],
                                        interval,
                                        year=year)

mindex = pd.MultiIndex.from_tuples([(i_swmx, i) for i in anom_merged.columns])

tmp = merged.copy()
tmp.columns = mindex
all_stations = pd.concat([tmp, all_stations], axis=1)

Average Conditions

The plots below shows the time-series of average soil moisture

Time-series

from model_comparison import plot_only_ts
import matplotlib.pyplot as plt
import numpy as np

avg_stations = all_stations.groupby(level=1, axis=1).mean()
anom_merged = plot_only_ts(avg_stations[
                               ['SWMX', 'PREVAH', 'ERA5', 'ERA5l', 'LISFLOOD', 'UERRA']
                           ], '', 'Average', [2012, 2018], anomalies=True, monthly=True)

Scatterplots

# create scatterplot
fig, axs = plt.subplots(2, 3, figsize=(15, 10))
collist = ['PREVAH', 'ERA5', 'ERA5l', 'LISFLOOD', 'UERRA']

scattermerged = anom_merged

# scattermerged[scattermerged.isnull().any(axis=1)] = np.nan
scattermerged = scattermerged.dropna(axis=0)

pearsonr = scattermerged.corr()

for i in range(5):
    if anomalies:
        plotlims = [-2.5, 2.5]
    else:
        plotlims = [0.1, 0.7]
    scattermerged.plot.scatter(x='SWMX', y=collist[i], ax=axs[np.unravel_index(i, (2, 3))],
                               title='SwissSMEX vs ' + collist[i],
                               xlim=plotlims, ylim=plotlims)
    # calculate rmse
    rmse = ((scattermerged[collist[i]] - scattermerged['SWMX']) ** 2).mean() ** .5
    axs[np.unravel_index(i, (2, 3))].text(0.1, 0.1,
                                          "R: " + "{:10.3f}".format(pearsonr['SWMX'][collist[i]]) + '\n' +
                                          "RMSE: " + "{:10.3f}".format(rmse),
                                          transform=axs[np.unravel_index(i, (2, 3))].transAxes,
                                          size='large')

fig.tight_layout()

Drought Sensitivity

Next we want to evaluate the ability of different soil moisture models to detect drought. For this purpose we classify drought based on three SMA thresholds (<-1, <-1.5, <-2) and evaluate how well the models can capture drought events. We used the following metrics to quantify the classification accuracy:

• true-positive-rate (TPR), $TPR=TP/P$
• false-positive-rate (FPR), $FPR=FP/N$
• true-negative-rate (TNR), $TNR=TN/N$
• false-negative-rate (FNR), $FNR=FN/P$
• accuracy (ACC), $ACC=(TP+TN)/(P+N)$

Where P is the true number of "positive" classifications (i.e. drought) and N the number of "negative" classifications (i.e. no drought).

SMA < -1

from model_comparison import sma_threshold

sma_threshold(-1, anom_merged)

	ERA5	ERA5l	LISFLOOD	PREVAH	UERRA
TPR	0.555556	0.552426	0.657277	0.589984	0.586854
FNR	0.444444	0.447574	0.342723	0.410016	0.413146
FPR	0.096677	0.106495	0.095921	0.104607	0.098943
TNR	0.903323	0.893505	0.904079	0.895393	0.901057
ACC	0.835716	0.827198	0.8561	0.836021	0.839976

SMA < -1.5

sma_threshold(-1.5, anom_merged)

	ERA5	ERA5l	LISFLOOD	PREVAH	UERRA
TPR	0.348659	0.348659	0.45977	0.425287	0.417625
FNR	0.651341	0.651341	0.54023	0.574713	0.582375
FPR	0.079974	0.080304	0.071381	0.080304	0.030403
TNR	0.920026	0.919696	0.928619	0.919696	0.969597
ACC	0.874658	0.874354	0.89139	0.880438	0.925768

SMA < -2.0

sma_threshold(-2, anom_merged)

	ERA5	ERA5l	LISFLOOD	PREVAH	UERRA
TPR	0.088608	0.101266	0.088608	0.151899	0.063291
FNR	0.911392	0.898734	0.911392	0.848101	0.936709
FPR	0.029925	0.036783	0.031172	0.044888	0.004988
TNR	0.970075	0.963217	0.968828	0.955112	0.995012
ACC	0.94889	0.942501	0.947673	0.935808	0.972619

Observations

• LISFLOOD seems to perform best overall (outperforming even PREVAH)
• The uncertainty of modelled soil moisture seems to be higher during winter (larger discrepancies are visible in the time-series plots during this period)
• The coarse resolution models perform similar
• Scatterplots show a clear linear relationship between in-situ measurements and modelled SMAs
• The level of noise is high for all models